xen/mmu: Use Xen specific TLB flush instead of the generic one.
authorKonrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Wed, 31 Oct 2012 16:38:31 +0000 (12:38 -0400)
committerKonrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Wed, 31 Oct 2012 16:38:31 +0000 (12:38 -0400)
commit95a7d76897c1e7243d4137037c66d15cbf2cce76
treed70ce3f79c8c5f7824ff277b2ab424ab975fdf2c
parentc8d258a7b2b6633b79cad4464e8dfb3dc328d897
xen/mmu: Use Xen specific TLB flush instead of the generic one.

As Mukesh explained it, the MMUEXT_TLB_FLUSH_ALL allows the
hypervisor to do a TLB flush on all active vCPUs. If instead
we were using the generic one (which ends up being xen_flush_tlb)
we end up making the MMUEXT_TLB_FLUSH_LOCAL hypercall. But
before we make that hypercall the kernel will IPI all of the
vCPUs (even those that were asleep from the hypervisor
perspective). The end result is that we needlessly wake them
up and do a TLB flush when we can just let the hypervisor
do it correctly.

This patch gives around 50% speed improvement when migrating
idle guest's from one host to another.

Oracle-bug: 14630170

CC: stable@vger.kernel.org
Tested-by:  Jingjie Jiang <jingjie.jiang@oracle.com>
Suggested-by:  Mukesh Rathor <mukesh.rathor@oracle.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
arch/x86/xen/mmu.c
include/trace/events/xen.h